The MetaX C600 144GB is the flagship AI accelerator from the Chinese company MetaX Integrated Circuits (Shanghai). Founded in September 2020 by former AMD and NVIDIA engineers, the company went public on the SSE STAR Market (688802) in December 2025.
The accelerator is positioned as a «universal AI GPU for both training and inference.» It is the only Chinese GPU with memory capacity on par with the NVIDIA H200 (141 GB).
The MetaX C600 144GB is the ideal tool for the new era of AI. It allows models of 70B+ scale in FP16 and models exceeding 120B+ in energy-efficient FP8 to be hosted on a single card — a feat previously only possible in multi-module configurations. Thus, the MetaX C600 doesn’t just aim to make tasks faster; it aims to make them possible — moving the inference of ultra-large models from laboratories into industrial production on standard server hardware. This transforms it from a simple accelerator into a strategic asset for any company working with cutting-edge generative models.
Technical Specifications: MetaX C600 144GB
Interface: PCIe 5.0 x16 (Maximum bandwidth)
Form Factor: PCIe Card / OAM (Two variants)
Architecture: Proprietary GPU IP (MetaX) (Fully independent development)
Process Node: Domestic (SMIC/partners) (Fully domestic)
Memory: 144 GB HBM3e
Memory Bandwidth: 3–4 TB/s (On par with NVIDIA H200 141GB)
Interconnect: MetaXLink (Multi-GPU scaling)
Precision Support: FP32, FP16, BF16, INT8, FP8 (Including FP8 for AI)
ECC/RAS: Yes (Enterprise-level reliability)
TDP: 300–400 W (For HBM3e + FP8 accelerators)
Comparison with Analogues:
Feature MetaX C600 NVIDIA H200 NVIDIA H100 Moore Threads S5000 Huawei Ascend 910C
Memory 144 GB HBM3e 141 GB HBM3e 80 GB HBM3 80 GB HBM 48 GB HBM2e
FP8 Yes Yes No Yes Yes
Interconnect MetaXLink NVLink 4.0 NVLink 4.0 MTLink HCCS
Process SMIC (Domestic) TSMC 4nm TSMC 4nm TSMC 12nm SMIC
TDP 300–400W 700W 700W ~350W ~400W
Architecture Features: MetaX C600 144GB
GPU Architecture: The C600 is built on a proprietary GPU IP architecture with a fully independent instruction set.
HBM3e 144 GB: On par with the H200. Allows 70B+ models in FP16 on a single card.
FP8: Hardware acceleration. Reduces power consumption by 30–50%, enabling 120B+ models in FP8.
MetaXLink: Proprietary multi-GPU interconnect for supernodes.
ECC/RAS: Error correction, enterprise-level reliability.
Multi-Precision: FP32, FP16, BF16, INT8, FP8.
Form Factor: PCIe Card / OAM.
Memory: 144 GB HBM3e — a record among Chinese GPUs.
Process Node: 7nm. Claimed to be a fully domestic supply chain — SMIC and partners.
Software Stack: MXMACA
CUDA Compatibility: CUDA code translation.
Frameworks: PyTorch, TensorFlow, vLLM, TGI, ONNX Runtime.
500+ AI models verified.
4,000+ open-source projects tested.
~50,000 developers in the community.
Preliminary Performance Evaluation
144 GB Memory: The only Chinese GPU with competitive capacity on par with the H200.
FP8 Support: Critical for efficient inference of large models.
Fully Domestic Supply Chain: Lack of dependence on TSMC reduces sanction risks.
Peak Performance: Estimated to be between the H100 and H200.
Software Stack Maturity: MXMACA is mature, but there is still a gap compared to CUDA.
Performance Estimates:
Inference vs. H100: 80–90% (CITIC Securities)
Cost vs. H100: 40–55% (CITIC Securities)
Inference DeepSeek-V3: 75–85% of H100 (ChaiTex internal tests)
Inference Qwen-72B: 75–85% of H100 (ChaiTex internal tests)
Inference Llama-3-70B: 75–85% of H100 (ChaiTex internal tests)
Fine-tuning LoRA 70B: Runs stably on 8 cards (ChaiTex internal tests)
Use Cases
LLM Inference 70B–200B+: One 144 GB card hosts the model entirely in FP16. In an 8x configuration = 1152 GB — for 400B+ models in FP8.
Fine-tuning / LoRA: 144 GB allows fine-tuning models up to 30B on a single card. On 8x — up to 70B.
Training from scratch: Up to 7B–13B on one card; up to 30B on an 8x configuration.
RAG Systems: High memory capacity allows keeping embedding models and vector DBs in GPU memory.
Multimodal Models: Processing images + text with large context windows.
AI for Science: Molecular dynamics, materials science, fluid dynamics.
Government / Regulated Tasks: Fully domestic supply chain meets import substitution requirements.
Pre-training foundation models: Requires clusters of hundreds/thousands of cards.
Edge inference: Not suitable due to high power consumption and form factor.
Testing Information
Official MetaX Testing:
500+ AI models verified on C600 via MXMACA.
4,000+ open-source projects tested.
Day-0 adaptation for DeepSeek V4 (April 2026).
Day-0 adaptation for Qwen3 (May 2026).
Independent Benchmarks:
MLPerf: No independent results yet (June 2026). MetaX has declared intent to participate in MLPerf Inference 2026.
CITIC Securities (May 2026): Estimates performance at 80–90% of H100 at 40–55% of the cost.
Real-world Deployment Tests (ChaiTex):
Model: DeepSeek-V3; Framework: vLLM + MXMACA; Result: 75–85% of H100 throughput.
Model: Qwen-72B-Chat; Framework: vLLM + MXMACA; Result: 75–85% of H100 throughput.
Model: Llama-3-70B-Instruct; Framework: vLLM + MXMACA; Result: 75–85% of H100 throughput.
Model: Fine-tuning LoRA 70B; Framework: PyTorch + MXMACA; Result: Stable on 8x C600.
Model: Embedding (BGE-large); Framework: ONNX Runtime + MXMACA; Result: Full memory utilization.
Note: All ChaiTex tests were conducted on an 8x C600 server configuration.
Deployments
China:
Muxing-1 (沐星一号): Server with 8x C600. Presented May 25, 2026, in Ningxia. Partnership with Ningxia Radio & TV, China Unicom Ningxia, Kehua Data Energy.
Government AI Platforms: Provincial «智算-centers» under the «East Data, West Compute» initiative. Requirement for «fully domestic» equipment.
Financial Institutions: Risk modeling, fraud detection, algorithmic trading. Large banks — internal LLMs.
Carrier 智算-centers: China Mobile, China Unicom — cloud AI infrastructures.
Integrators: Baixin Information, Xinghan Intelligent.
Russia (via ChaiTex):
Recommended for enterprise inference, fine-tuning, RAG.
GPU rental in Russian and Chinese data centers.
Pilot projects with major integrators (NDA).
Server configurations available for delivery.
Rating and Evaluation
Raw Compute Performance: ★★★★☆ — Estimated between H100 and H200.
Memory Capacity: ★★★★★ — 144 GB HBM3e is the best among Chinese GPUs.
Software Stack: ★★★★☆ — MXMACA is mature, but there is a gap compared to CUDA.
Reliability / ECC: ★★★★★ — Built-in ECC/RAS.
Scalability: ★★★★☆ — MetaXLink supports 8-GPU supernodes.
Energy Efficiency: ★★★★☆ — HBM3e + FP8, TDP lower than H100/H200.
Price / TCO: ★★★★★ — 40–55% of H200 cost for 80–90% of performance.
Sovereignty / Sanction Risk: ★★★★★ — SMIC production.
Overall Rating: 4.6 / 5
The accelerator is available for pre-order in the Chaitex catalog.
#1 among Chinese GPUs for inference tasks of 70B+ parameters.
#1 in memory capacity among all Chinese GPUs.
#2 for fine-tuning — trails S5000 in compute, but surpasses in memory.
#3 for training from scratch — trails Huawei Ascend 950DT and Moore Threads S5000.
Energy Efficiency
HBM3e: High bandwidth with lower power consumption.
FP8: Reduces power consumption by 30–50% vs FP16.
MetaXLink: Energy-efficient interconnect without PCIe overhead.
SMIC Process: TDP of 300–400W is lower than H100/H200 (700W).
Power Consumption Comparison:
MetaX C600: TDP 300–400W; Power per GB memory: 2.1–2.8 W/GB.
NVIDIA H200: TDP 700W; Power per GB memory: 5.0 W/GB.
NVIDIA H100: TDP 700W; Power per GB memory: 8.75 W/GB.
Moore Threads S5000: TDP ~350W; Power per GB memory: 4.4 W/GB.
Comparison:
The C600 consumes roughly half the energy for a comparable memory capacity.
Support for air cooling is an advantage for TCO.
Conclusions and Summary
The MetaX C600 144GB is the first Chinese GPU that competes with NVIDIA not as a «choice of last resort,» but on real technical merits.
Key Advantages:
Record Memory Capacity: 144 GB, on par with the H200.
Fully Domestic Supply Chain: Minimal sanction risks.
Low Power Consumption: 300–400W vs 700W for H100/H200.
Mature Software: MXMACA with 500+ models.
Who this card is for:
Enterprise customers with LLM inference 70B–200B+.
Government organizations with import substitution requirements.
Financial and medical institutions requiring data sovereignty.
Companies starting AI transformation — CUDA compatibility lowers the migration barrier.
If the task is production inference of large LLMs (DeepSeek, Qwen, Llama 70B+) in a closed environment with sovereignty requirements — the MetaX C600 144GB is the optimal choice in the Chinese market.
The MetaX C600 marks a qualitative shift from import substitution to full-fledged competition. It is not just a «fallback alternative,» but a product that forces global players to reconsider their strategy. The appearance of a GPU with 144 GB of memory, produced via a fully domestic supply chain, radically reduces sanction risks for Chinese corporations and state structures, making them independent of US export controls. For the rest of the world, this is a signal that the monopoly on advanced AI compute is broken, and customers will now evaluate suppliers not only on raw performance but also on factors like technological sovereignty, supply predictability, and Total Cost of Ownership (TCO).
The success of the C600, built on SMIC’s 7nm process, is a crucial validation event for the entire Chinese semiconductor industry. It proves that even without access to advanced 4nm and 3nm technologies, competitive products can be created by compensating for transistor density deficits with architectural innovations. In the short term, this threatens NVIDIA’s market share in the mid-range. In the long term — the launch of the C600 creates the necessary foundation for developing the next generation of chips (e.g., the expected C700-series), which will be aimed at direct competition with NVIDIA’s H100/H200 on all fronts, including raw compute, and will finally cement the new multipolar reality in the accelerator market.
Contact our experts to select the optimal hardware configuration and calculate rental pricing for your AI workloads.